History

Merge branch 'net_proc_perf'

Jia He says: ==================== Reduce cache miss for snmp_fold_field In a PowerPc server with large cpu number(160), besides commit a3a773726c9f ("net: Optimize snmp stat aggregation by walking all the percpu data at once"), I watched several other snmp_fold_field callsites which would cause high cache miss rate. test source code: ================ My simple test case, which read from the procfs items endlessly: /***********************************************************/ int main(int argc, char **argv) { int i; int fd = -1 ; int rdsize = 0; char buf[LINELEN+1]; buf[LINELEN] = 0; memset(buf,0,LINELEN); if(1 >= argc) { printf("file name empty\n"); return -1; } fd = open(argv[1], O_RDWR, 0644); if(0 > fd){ printf("open error\n"); return -2; } for(i=0;i<0xffffffff;i++) { while(0 < (rdsize = read(fd,buf,LINELEN))){ //nothing here } lseek(fd, 0, SEEK_SET); } close(fd); return 0; } /**********************************************************/ compile and run: ================ gcc test.c -o test perf stat -d -e cache-misses ./test /proc/net/snmp perf stat -d -e cache-misses ./test /proc/net/snmp6 perf stat -d -e cache-misses ./test /proc/net/sctp/snmp perf stat -d -e cache-misses ./test /proc/net/xfrm_stat before the patch set: ==================== Performance counter stats for 'system wide': 355911097 cache-misses [40.08%] 2356829300 L1-dcache-loads [60.04%] 355642645 L1-dcache-load-misses # 15.09% of all L1-dcache hits [60.02%] 346544541 LLC-loads [59.97%] 389763 LLC-load-misses # 0.11% of all LL-cache hits [40.02%] 6.245162638 seconds time elapsed After the patch set: =================== Performance counter stats for 'system wide': 194992476 cache-misses [40.03%] 6718051877 L1-dcache-loads [60.07%] 194871921 L1-dcache-load-misses # 2.90% of all L1-dcache hits [60.11%] 187632232 LLC-loads [60.04%] 464466 LLC-load-misses # 0.25% of all LL-cache hits [39.89%] 6.868422769 seconds time elapsed The cache-miss rate can be reduced from 15% to 2.9% changelog ========= v6: - correct v5 v5: - order local variables from longest to shortest line v4: - move memset into one block of if statement in snmp6_seq_show_item - remove the changes in netstat_seq_show considerred the stack usage is too large v3: - introduce generic interface (suggested by Marcelo Ricardo Leitner) - use max_t instead of self defined macro (suggested by David Miller) v2: - fix bug in udplite statistics. - snmp_seq_show is split into 2 parts ====================

David S. Miller committed 10 years ago

Commit bcdc6ef2 parents fa14035 6d4a7416 files · +129 −58Browse files

Unified · Split

6 files
  1. include/net/ip.h +23 −0
  2. net/ipv4/proc.c +63 −39
  3. net/ipv6/addrconf.c +6 −6
  4. net/ipv6/proc.c +22 −8
  5. net/sctp/proc.c +7 −3
  6. net/xfrm/xfrm_proc.c +8 −2
include/net/ip.h+23−0
@@ -219,6 +219,29 @@ static inline u64 snmp_fold_field64(void __percpu *mib, int offt, size_t syncp_o · 218 unmodified lines
219219 }
220220 #endif
221221
222+#define snmp_get_cpu_field64_batch(buff64, stats_list, mib_statistic, offset) \
223+{ \
224+ int i, c; \
225+ for_each_possible_cpu(c) { \
226+ for (i = 0; stats_list[i].name; i++) \
227+ buff64[i] += snmp_get_cpu_field64( \
228+ mib_statistic, \
229+ c, stats_list[i].entry, \
230+ offset); \
231+ } \
232+}
233+
234+#define snmp_get_cpu_field_batch(buff, stats_list, mib_statistic) \
235+{ \
236+ int i, c; \
237+ for_each_possible_cpu(c) { \
238+ for (i = 0; stats_list[i].name; i++) \
239+ buff[i] += snmp_get_cpu_field( \
240+ mib_statistic, \
241+ c, stats_list[i].entry); \
242+ } \
243+}
244+
222245 void inet_get_local_port_range(struct net *net, int *low, int *high);
223246
224247 #ifdef CONFIG_SYSCTL
net/ipv4/proc.c+63−39
@@ -46,6 +46,8 @@ · 45 unmodified lines
4646 #include <net/sock.h>
4747 #include <net/raw.h>
4848
49+#define TCPUDP_MIB_MAX max_t(u32, UDP_MIB_MAX, TCP_MIB_MAX)
50+
4951 /*
5052 * Report socket allocation statistics [mea@utu.fi]
5153 */
@@ -356,22 +358,22 @@ static void icmp_put(struct seq_file *seq) · 304 unmodified lines
356358 atomic_long_t *ptr = net->mib.icmpmsg_statistics->mibs;
357359
358360 seq_puts(seq, "\nIcmp: InMsgs InErrors InCsumErrors");
359− for (i = 0; icmpmibmap[i].name != NULL; i++)
361+ for (i = 0; icmpmibmap[i].name; i++)
360362 seq_printf(seq, " In%s", icmpmibmap[i].name);
361363 seq_puts(seq, " OutMsgs OutErrors");
362− for (i = 0; icmpmibmap[i].name != NULL; i++)
364+ for (i = 0; icmpmibmap[i].name; i++)
363365 seq_printf(seq, " Out%s", icmpmibmap[i].name);
364366 seq_printf(seq, "\nIcmp: %lu %lu %lu",
365367 snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_INMSGS),
366368 snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_INERRORS),
367369 snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_CSUMERRORS));
368− for (i = 0; icmpmibmap[i].name != NULL; i++)
370+ for (i = 0; icmpmibmap[i].name; i++)
369371 seq_printf(seq, " %lu",
370372 atomic_long_read(ptr + icmpmibmap[i].index));
371373 seq_printf(seq, " %lu %lu",
372374 snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_OUTMSGS),
373375 snmp_fold_field(net->mib.icmp_statistics, ICMP_MIB_OUTERRORS));
374− for (i = 0; icmpmibmap[i].name != NULL; i++)
376+ for (i = 0; icmpmibmap[i].name; i++)
375377 seq_printf(seq, " %lu",
376378 atomic_long_read(ptr + (icmpmibmap[i].index | 0x100)));
377379 }
@@ -379,14 +381,16 @@ static void icmp_put(struct seq_file *seq) · 1 unmodified line
379381 /*
380382 * Called from the PROCfs module. This outputs /proc/net/snmp.
381383 */
382−static int snmp_seq_show(struct seq_file *seq, void *v)
384+static int snmp_seq_show_ipstats(struct seq_file *seq, void *v)
383385 {
384− int i;
385386 struct net *net = seq->private;
387+ u64 buff64[IPSTATS_MIB_MAX];
388+ int i;
386389
387− seq_puts(seq, "Ip: Forwarding DefaultTTL");
390+ memset(buff64, 0, IPSTATS_MIB_MAX * sizeof(u64));
388391
389− for (i = 0; snmp4_ipstats_list[i].name != NULL; i++)
392+ seq_puts(seq, "Ip: Forwarding DefaultTTL");
393+ for (i = 0; snmp4_ipstats_list[i].name; i++)
390394 seq_printf(seq, " %s", snmp4_ipstats_list[i].name);
391395
392396 seq_printf(seq, "\nIp: %d %d",
@@ -394,57 +398,77 @@ static int snmp_seq_show(struct seq_file *seq, void *v) · 1 unmodified line
394398 net->ipv4.sysctl_ip_default_ttl);
395399
396400 BUILD_BUG_ON(offsetof(struct ipstats_mib, mibs) != 0);
397− for (i = 0; snmp4_ipstats_list[i].name != NULL; i++)
398− seq_printf(seq, " %llu",
399− snmp_fold_field64(net->mib.ip_statistics,
400− snmp4_ipstats_list[i].entry,
401− offsetof(struct ipstats_mib, syncp)));
401+ snmp_get_cpu_field64_batch(buff64, snmp4_ipstats_list,
402+ net->mib.ip_statistics,
403+ offsetof(struct ipstats_mib, syncp));
404+ for (i = 0; snmp4_ipstats_list[i].name; i++)
405+ seq_printf(seq, " %llu", buff64[i]);
402406
403− icmp_put(seq); /* RFC 2011 compatibility */
404− icmpmsg_put(seq);
407+ return 0;
408+}
409+
410+static int snmp_seq_show_tcp_udp(struct seq_file *seq, void *v)
411+{
412+ unsigned long buff[TCPUDP_MIB_MAX];
413+ struct net *net = seq->private;
414+ int i;
415+
416+ memset(buff, 0, TCPUDP_MIB_MAX * sizeof(unsigned long));
405417
406418 seq_puts(seq, "\nTcp:");
407− for (i = 0; snmp4_tcp_list[i].name != NULL; i++)
419+ for (i = 0; snmp4_tcp_list[i].name; i++)
408420 seq_printf(seq, " %s", snmp4_tcp_list[i].name);
409421
410422 seq_puts(seq, "\nTcp:");
411− for (i = 0; snmp4_tcp_list[i].name != NULL; i++) {
423+ snmp_get_cpu_field_batch(buff, snmp4_tcp_list,
424+ net->mib.tcp_statistics);
425+ for (i = 0; snmp4_tcp_list[i].name; i++) {
412426 /* MaxConn field is signed, RFC 2012 */
413427 if (snmp4_tcp_list[i].entry == TCP_MIB_MAXCONN)
414− seq_printf(seq, " %ld",
415− snmp_fold_field(net->mib.tcp_statistics,
416− snmp4_tcp_list[i].entry));
428+ seq_printf(seq, " %ld", buff[i]);
417429 else
418− seq_printf(seq, " %lu",
419− snmp_fold_field(net->mib.tcp_statistics,
420− snmp4_tcp_list[i].entry));
430+ seq_printf(seq, " %lu", buff[i]);
421431 }
422432
433+ memset(buff, 0, TCPUDP_MIB_MAX * sizeof(unsigned long));
434+
435+ snmp_get_cpu_field_batch(buff, snmp4_udp_list,
436+ net->mib.udp_statistics);
423437 seq_puts(seq, "\nUdp:");
424− for (i = 0; snmp4_udp_list[i].name != NULL; i++)
438+ for (i = 0; snmp4_udp_list[i].name; i++)
425439 seq_printf(seq, " %s", snmp4_udp_list[i].name);
426−
427440 seq_puts(seq, "\nUdp:");
428− for (i = 0; snmp4_udp_list[i].name != NULL; i++)
429− seq_printf(seq, " %lu",
430− snmp_fold_field(net->mib.udp_statistics,
431− snmp4_udp_list[i].entry));
441+ for (i = 0; snmp4_udp_list[i].name; i++)
442+ seq_printf(seq, " %lu", buff[i]);
443+
444+ memset(buff, 0, TCPUDP_MIB_MAX * sizeof(unsigned long));
432445
433446 /* the UDP and UDP-Lite MIBs are the same */
434447 seq_puts(seq, "\nUdpLite:");
435− for (i = 0; snmp4_udp_list[i].name != NULL; i++)
448+ snmp_get_cpu_field_batch(buff, snmp4_udp_list,
449+ net->mib.udplite_statistics);
450+ for (i = 0; snmp4_udp_list[i].name; i++)
436451 seq_printf(seq, " %s", snmp4_udp_list[i].name);
437−
438452 seq_puts(seq, "\nUdpLite:");
439− for (i = 0; snmp4_udp_list[i].name != NULL; i++)
440− seq_printf(seq, " %lu",
441− snmp_fold_field(net->mib.udplite_statistics,
442− snmp4_udp_list[i].entry));
453+ for (i = 0; snmp4_udp_list[i].name; i++)
454+ seq_printf(seq, " %lu", buff[i]);
443455
444456 seq_putc(seq, '\n');
445457 return 0;
446458 }
447459
460+static int snmp_seq_show(struct seq_file *seq, void *v)
461+{
462+ snmp_seq_show_ipstats(seq, v);
463+
464+ icmp_put(seq); /* RFC 2011 compatibility */
465+ icmpmsg_put(seq);
466+
467+ snmp_seq_show_tcp_udp(seq, v);
468+
469+ return 0;
470+}
471+
448472 static int snmp_seq_open(struct inode *inode, struct file *file)
449473 {
450474 return single_open_net(inode, file, snmp_seq_show);
@@ -469,21 +493,21 @@ static int netstat_seq_show(struct seq_file *seq, void *v) · 18 unmodified lines
469493 struct net *net = seq->private;
470494
471495 seq_puts(seq, "TcpExt:");
472− for (i = 0; snmp4_net_list[i].name != NULL; i++)
496+ for (i = 0; snmp4_net_list[i].name; i++)
473497 seq_printf(seq, " %s", snmp4_net_list[i].name);
474498
475499 seq_puts(seq, "\nTcpExt:");
476− for (i = 0; snmp4_net_list[i].name != NULL; i++)
500+ for (i = 0; snmp4_net_list[i].name; i++)
477501 seq_printf(seq, " %lu",
478502 snmp_fold_field(net->mib.net_statistics,
479503 snmp4_net_list[i].entry));
480504
481505 seq_puts(seq, "\nIpExt:");
482− for (i = 0; snmp4_ipextstats_list[i].name != NULL; i++)
506+ for (i = 0; snmp4_ipextstats_list[i].name; i++)
483507 seq_printf(seq, " %s", snmp4_ipextstats_list[i].name);
484508
485509 seq_puts(seq, "\nIpExt:");
486− for (i = 0; snmp4_ipextstats_list[i].name != NULL; i++)
510+ for (i = 0; snmp4_ipextstats_list[i].name; i++)
487511 seq_printf(seq, " %llu",
488512 snmp_fold_field64(net->mib.ip_statistics,
489513 snmp4_ipextstats_list[i].entry,
net/ipv6/addrconf.c+6−6
@@ -4961,18 +4961,18 @@ static inline size_t inet6_if_nlmsg_size(void) · 4960 unmodified lines
49614961 }
49624962
49634963 static inline void __snmp6_fill_statsdev(u64 *stats, atomic_long_t *mib,
4964− int items, int bytes)
4964+ int bytes)
49654965 {
49664966 int i;
4967− int pad = bytes - sizeof(u64) * items;
4967+ int pad = bytes - sizeof(u64) * ICMP6_MIB_MAX;
49684968 BUG_ON(pad < 0);
49694969
49704970 /* Use put_unaligned() because stats may not be aligned for u64. */
4971− put_unaligned(items, &stats[0]);
4972− for (i = 1; i < items; i++)
4971+ put_unaligned(ICMP6_MIB_MAX, &stats[0]);
4972+ for (i = 1; i < ICMP6_MIB_MAX; i++)
49734973 put_unaligned(atomic_long_read(&mib[i]), &stats[i]);
49744974
4975− memset(&stats[items], 0, pad);
4975+ memset(&stats[ICMP6_MIB_MAX], 0, pad);
49764976 }
49774977
49784978 static inline void __snmp6_fill_stats64(u64 *stats, void __percpu *mib,
@@ -5005,7 +5005,7 @@ static void snmp6_fill_stats(u64 *stats, struct inet6_dev *idev, int attrtype, · 26 unmodified lines
50055005 offsetof(struct ipstats_mib, syncp));
50065006 break;
50075007 case IFLA_INET6_ICMP6STATS:
5008− __snmp6_fill_statsdev(stats, idev->stats.icmpv6dev->mibs, ICMP6_MIB_MAX, bytes);
5008+ __snmp6_fill_statsdev(stats, idev->stats.icmpv6dev->mibs, bytes);
50095009 break;
50105010 }
50115011 }
net/ipv6/proc.c+22−8
@@ -30,6 +30,11 @@ · 29 unmodified lines
3030 #include <net/transp_v6.h>
3131 #include <net/ipv6.h>
3232
33+#define MAX4(a, b, c, d) \
34+ max_t(u32, max_t(u32, a, b), max_t(u32, c, d))
35+#define SNMP_MIB_MAX MAX4(UDP_MIB_MAX, TCP_MIB_MAX, \
36+ IPSTATS_MIB_MAX, ICMP_MIB_MAX)
37+
3338 static int sockstat6_seq_show(struct seq_file *seq, void *v)
3439 {
3540 struct net *net = seq->private;
@@ -191,25 +196,34 @@ static void snmp6_seq_show_item(struct seq_file *seq, void __percpu *pcpumib, · 155 unmodified lines
191196 atomic_long_t *smib,
192197 const struct snmp_mib *itemlist)
193198 {
199+ unsigned long buff[SNMP_MIB_MAX];
194200 int i;
195− unsigned long val;
196201
197− for (i = 0; itemlist[i].name; i++) {
198− val = pcpumib ?
199− snmp_fold_field(pcpumib, itemlist[i].entry) :
200− atomic_long_read(smib + itemlist[i].entry);
201− seq_printf(seq, "%-32s\t%lu\n", itemlist[i].name, val);
202+ if (pcpumib) {
203+ memset(buff, 0, sizeof(unsigned long) * SNMP_MIB_MAX);
204+
205+ snmp_get_cpu_field_batch(buff, itemlist, pcpumib);
206+ for (i = 0; itemlist[i].name; i++)
207+ seq_printf(seq, "%-32s\t%lu\n",
208+ itemlist[i].name, buff[i]);
209+ } else {
210+ for (i = 0; itemlist[i].name; i++)
211+ seq_printf(seq, "%-32s\t%lu\n", itemlist[i].name,
212+ atomic_long_read(smib + itemlist[i].entry));
202213 }
203214 }
204215
205216 static void snmp6_seq_show_item64(struct seq_file *seq, void __percpu *mib,
206217 const struct snmp_mib *itemlist, size_t syncpoff)
207218 {
219+ u64 buff64[SNMP_MIB_MAX];
208220 int i;
209221
222+ memset(buff64, 0, sizeof(unsigned long) * SNMP_MIB_MAX);
223+
224+ snmp_get_cpu_field64_batch(buff64, itemlist, mib, syncpoff);
210225 for (i = 0; itemlist[i].name; i++)
211− seq_printf(seq, "%-32s\t%llu\n", itemlist[i].name,
212− snmp_fold_field64(mib, itemlist[i].entry, syncpoff));
226+ seq_printf(seq, "%-32s\t%llu\n", itemlist[i].name, buff64[i]);
213227 }
214228
215229 static int snmp6_seq_show(struct seq_file *seq, void *v)
net/sctp/proc.c+7−3
@@ -73,13 +73,17 @@ static const struct snmp_mib sctp_snmp_list[] = { · 72 unmodified lines
7373 /* Display sctp snmp mib statistics(/proc/net/sctp/snmp). */
7474 static int sctp_snmp_seq_show(struct seq_file *seq, void *v)
7575 {
76+ unsigned long buff[SCTP_MIB_MAX];
7677 struct net *net = seq->private;
7778 int i;
7879
79− for (i = 0; sctp_snmp_list[i].name != NULL; i++)
80+ memset(buff, 0, sizeof(unsigned long) * SCTP_MIB_MAX);
81+
82+ snmp_get_cpu_field_batch(buff, sctp_snmp_list,
83+ net->sctp.sctp_statistics);
84+ for (i = 0; sctp_snmp_list[i].name; i++)
8085 seq_printf(seq, "%-32s\t%ld\n", sctp_snmp_list[i].name,
81− snmp_fold_field(net->sctp.sctp_statistics,
82− sctp_snmp_list[i].entry));
86+ buff[i]);
8387
8488 return 0;
8589 }
net/xfrm/xfrm_proc.c+8−2
@@ -50,12 +50,18 @@ static const struct snmp_mib xfrm_mib_list[] = { · 49 unmodified lines
5050
5151 static int xfrm_statistics_seq_show(struct seq_file *seq, void *v)
5252 {
53+ unsigned long buff[LINUX_MIB_XFRMMAX];
5354 struct net *net = seq->private;
5455 int i;
56+
57+ memset(buff, 0, sizeof(unsigned long) * LINUX_MIB_XFRMMAX);
58+
59+ snmp_get_cpu_field_batch(buff, xfrm_mib_list,
60+ net->mib.xfrm_statistics);
5561 for (i = 0; xfrm_mib_list[i].name; i++)
5662 seq_printf(seq, "%-24s\t%lu\n", xfrm_mib_list[i].name,
57− snmp_fold_field(net->mib.xfrm_statistics,
58− xfrm_mib_list[i].entry));
63+ buff[i]);
64+
5965 return 0;
6066 }
6167

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t